OpenClaw 2026.3.22: GPT-5.4, Advanced Search Tools, and the New Plugin SDK

The OpenClaw 2026.3.22 release is here, and it’s one of the most significant updates to the ecosystem we've seen this year. Following the v2026.3.13 breakthrough in browser automation, this update focuses on maturing the agent framework, expanding its search capabilities, and refining how users interact with their autonomous systems.

From the transition to GPT-5.4 as the new default model to the long-awaited Plugin SDK, there’s a lot to unpack. But be warned: this release prioritizes security and modularity over backward compatibility. Let’s dive into everything new in OpenClaw 2026.3.22.

GPT-5.4: The New Intelligence Standard

For the first time since the v2026.3.12 release, OpenClaw has updated its default OpenAI setup model. The standard model is now openai/gpt-5.4, while the programming-optimized model remains openai-codex/gpt-5.4.

This shift provides agents with a significant boost in reasoning stability and tool-calling accuracy. To simplify management, OpenAI chat, image, TTS, transcription, and embedding defaults have been centralized into a single shard module, making it easier for users to track across complex agent fleets.

Furthermore, Per-Agent AI Defaults have been introduced. This allows users to set specific thinking and reasoning models for individual agents. If a manually provided model override is disallowed or fails, the system now automatically reverts to the agent's pre-configured default, preventing session hangs.

Production-Grade Secrets and PDF Intelligence

While the search tools are flashy, some of the most powerful updates are under the hood:

Advanced Secrets Management: OpenClaw now features a production-ready secrets system with SecretRef support for up to 64 credential targets. This makes it significantly safer to connect your agents to real-world accounts without leaking API keys in logs or configuration files.
First-Class PDF Support: Agents now have native PDF processing capabilities. This includes configurable page and size limits, as well as extraction fallbacks. This is a massive win for users relying on agents for document summarization or structured data extraction from clinical or financial reports.
Unified Outbound Adapter: A single sendPayload adapter now handles communication across Discord, Slack, WhatsApp, Zalo, and Telegram. This simplifies multi-channel agent logic and ensures consistent message delivery across the board.

A New Suite of Advanced Search Tools

Autonomous agents are only as good as the information they can access. OpenClaw 2026.3.22 introduces native integrations for three of the most powerful AI-first search and scraping tools:

Exa: Now includes native date filters and search-mode selection. This enables agents to find hyper-recent research papers or news within specific timeframes.
Tavily: Integrated with dedicated tavily_search and tavily_extract tools. Tavily is renowned for providing LLM-optimized search results, reducing token noise.
Firecrawl: The new firecrawl_search and firecrawl_scrape tools allow agents to convert any URL into clean, LLM-ready markdown instantly.

These additions make OpenClaw the most capable platform for automated research and data extraction. For more on setting up these integrations, check our Integration and Setup Guide.

The Modern Plugin SDK and ClawHub

The way we extend OpenClaw is changing, and this is where the "pain" begins for some. The legacy openclaw/extension-api has been officially retired without a compatibility shim.

Developers must now migrate to the new openclaw/plugin-sdk/* surface area. While this is a breaking change, it provides a far more modular and secure environment for everything from browser automation to voice-powered workflows.

Hand-in-hand with the SDK is the promotion of ClawHub as the primary installation source. When running openclaw plugins install <package>, the system will now prioritize ClawHub for npm-safe package names, using npm only as a fallback. This ensures that community-vetted, agent-specific plugins are always preferred.

UX Improvements: Clarifying Questions Mid-Session

One of the most frustrating parts of using an autonomous agent is when it makes a wrong assumption halfway through a 30-minute research task.

OpenClaw 2026.3.22 solves this by allowing agents to ask clarifying questions mid-session. These questions appear as non-blocking, dismissible in-session TUI responses. You can provide additional context or pick a preference without interrupting the agent's current thought process or context window.

Important Security Hardening and Migrations

As with any major update, some legacy paths are being cleared out to make room for more robust systems.

The Browser Automation Shift

The Legacy Chrome extension relay path has been removed. This includes driver: "extension" and browser.relayBindHost. If you rely on browser automation, you must migrate to existing-session / user mode.

Pro Tip: You can handle this migration automatically by running:

openclaw doctor --fix

New Default Permissions

New installations now default the tools.profile to "messaging". This means that filesystem and shell access (exec, web_fetch, etc.) are not granted by default. If your agent suddenly seems "limited" after a clean install, you will need to manually enable the "full" tools profile in your openclaw.json. This is a critical security measure to prevent unauthorized script execution.

If you encounter issues during this migration, refer to our Troubleshooting Guide or review the Security Best Practices for managing browser sessions.

Stability and Gateway Fixes

Docker & Sandbox: The removal of the extension relay does not affect Docker, headless, or sandboxed environments, which remain the recommended path for production-grade deployments.
macOS and Windows: This release fixes several environment-specific bugs related to service command reuse and exec approvals. macOS exec approvals now correctly respect per-agent settings.

OpenClaw 2026.3.22 is a major leap toward the production-grade reliability of autonomous agents. You can pull the latest update and read the full changelog from the official GitHub repository.

Ready to deploy? Follow our Step-by-Step Deployment Guide to get started with the latest version today.